Boston Dynamics Unveils Four-Fingered Atlas Hand
Boston Dynamics has completely revamped the hand of its human-like robot 'Atlas.' On Monday (local time), the Hyundai Motor Group subsidiary unveiled for the first time a next-generation robotic hand with a four-fingered structure including a thumb, and released a video showing an Atlas equipped with the new hands performing a variety of manipulation tasks, including drilling. Robotic hands are a core component of humanoids, enabling them to use the same workspaces and tools as humans. Unlike previous demos that focused on walking and running, this demonstration drew attention for directly showcasing the continuous, complex tasks demanded in industrial settings.
In the video, Atlas picks up a drill bit with its new hand, attaches it to a drill, and bores a hole into a piece of wood. Since the task requires gripping the handle while also pressing a button, it demands precise control over each finger's strength and movement. Atlas was also shown gripping a small nut with its fingertips, twisting it to tighten it. Other maneuvers included twirling a drumstick between its fingers and rolling two golf balls simultaneously within one hand. These tasks require continuously adjusting the contact and force between fingers without dropping the objects, proving that the hand can adapt its finger movements to objects of different shapes and materials.
Boston Dynamics explained that after analyzing industrial robot use cases and weighing factors such as design complexity, power consumption, cost, durability, and ease of repair, it chose a four-fingered structure rather than five. The team explored the optimal design through experiments, simulations, and 3D-printed mockups, and even had the entire development team work a full day with their pinky fingers immobilized, concluding that losing one finger would not significantly hinder performance.
From the early stages of development, Boston Dynamics employed a sim-to-real approach, first training the robot in virtual environments and then applying the results to the physical robot. By running reinforcement learning with varied motor torque, friction, object shapes, and disturbances occurring during tasks, the company continues to enhance manipulation capabilities that can handle a wide range of conditions.
The Atlas development model is set to be deployed starting in 2028 at Hyundai Motor Group's Metaplant America (HMGMA) in Georgia and Kia's Georgia plant in the United States.
